Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
20| Simon Kirby — Language Evolution & Emergence of Structure
193 Plays1 year ago
Language is the ultimate Lego. With it, we can take simple elements and construct them into an edifice of meaning. Its power is not only in mapping signs to concepts but in that individual words can be composed into larger structures.
How did this systematicity arise in language?
Simon Kirby is the head of Linguistics and English Language at The University of Edinburgh and one of the founders of the Centre for Langauge Evolution and Change. Over several decades he and his collaborators have run many elegant experiments that show that this property of language emerges inexorably as a system of communication is passed from generation to generation.
Experiments with computer simulations, humans, and even baboons demonstrate that as a language is learned mistakes are made - much like the mutations in genes. Crucially, the mistakes that better match the language to the structure of the world (as conceived by the learner) are the ones that are most likely to be passed on.
Links
- Simon's website with art, music, and talks on language evolution
- Simon's academic homepage
- Simon on X
- Multiverses Podcast home
Outline
(00:00) Introduction
(2:45) What makes language special?
(5:30) Language extends our biological bounds
(7:55) Language makes culture, culture makes language
(9:30) John Searle: world to word and word to world
(13:30) Compositionality: the expressivity of language is based on its Lego-like combinations
(16:30) Could unique genes explain the fact of language compositionality?
(17:20) … Not fully, though they might make our brains able to support compositional language
(18:20) Using simulations to model language learning and search for the emergence of structure
(19:35) Compositionality emerges from the transmission of representations across generations
(20:18) The learners need to make mistakes, but not random mistakes
(21:35) Just like biological evolution, we need variation
(27:00) When, by chance, linguistic features echo the structure of the world these are more memorable
(33:45) Language experiments with humans (Hannah Cornish)
(36:32) Sign language experiments in the lab (Yasamin Motamedi)
(38:45) Spontaneous emergence of sign language in populations
(41:18) Communication is key to making language efficient, while transmission gives structure
(47:10) Without intentional design these processes produce optimized systems
(50:39) We need to perceive similarity in states of the world for linguistic structure to emerge
(57:05) Why isn’t language ubiquitous in nature …
(58:00) … why do only humans have cultural transmissions
(59:56) Over-imitation: Victoria Horner & Andrew Whiten, humans love to copy each other
(1:06:00) Is language a spandrel?
(1:07:10) How much of language is about information transfer? Partner-swapping conversations (Gareth Roberts)
(1:08:49) Language learning = play?
(1:12:25) Iterated learning experiments with baboons (& Tetris!)
(1:17:50) Endogenous rewards for copying
(1:20:30) Art as another angle on the same problems
Recommended
Transcript
Introduction to Human Language and Uniqueness
00:00:00
Speaker
We humans have some pretty cool super pairs. We can pose our finger and thumb. We can strike matches and make fires. But of all the things we can do, it's perhaps our ability to create, manipulate, pass around, and store ideas using language that most marks us out from other species. Where do we get this? Well, we all learn a language when we're growing up. But how did that tool come about in the first place? Where does language come from?
Introducing Professor Simon Kirby
00:00:25
Speaker
I guess this week is Professor Simon Kirby. He's the head of linguistics and English language at the University of Edinburgh, where he's also one of the founders for the Centre for Language Evolution and Change.
Language Evolution Experiments and Systematicity
00:00:35
Speaker
And it's this that we'll be talking about. Simon will take us through a wonderfully elegant set of experiments that he's run over the last decades with many colleagues. In these experiments,
00:00:47
Speaker
agents, they could be computer programs, they could be people, they could even be baboons. Agents are taught a mapping of words, of sounds, symbols, or science to concepts. But the initial mapping that they're given is completely random, it's not systematic at all. There's no way of putting things together.
00:01:12
Speaker
Incredibly, what they show is that as this language, this kind of proto-language, if you like, is passed down through generation to generation, so it's taught to first generation, they teach it to a second generation, and so on in what's called iterated learning experiments, which Simon and his team have really pioneered. As this happens, systematicity emerges. Incredibly, the fallibility of learners, the mistakes that are made,
00:01:40
Speaker
as the language it passes from generation to generation, don't just add noise, they add up to something.
Art and Linguistics: A Brief Intersection
00:01:47
Speaker
Just like with biological evolution, there seems to be a direction that actually doesn't come from design, but just from simple constraints. So this is an incredible founding, and I really enjoy going over the wonderfully imaginative experiments that have been run in this field.
00:02:08
Speaker
We end up talking a little bit about art, and I should mention that Simon really underplays his achievements here.
Podcast Introduction and Language as a Human Trait
00:02:13
Speaker
He actually has won a BAFTA for one of his pieces. If you enjoy this conversation, I'd strongly encourage you to Google Simon Kirby, look at his artworks, listen to the album that he's just released, and enjoy these beautiful experiments that we discuss.
00:02:29
Speaker
And with that, I'm James Robinson. You're listening to Multiverses. Simon Kirby, thanks for joining me. You're welcome. Happy to be here. What's special about language?
00:02:57
Speaker
That's a great place to start, the question of why would we study this thing at all, really? I mean, one answer to that is that it is the thing probably above anything else that marks us out as human. It's, I would say, it's our species' best trick. I mean, all species have something special. And for us, it's language, I think.
00:03:26
Speaker
So if we are to understand what it means to be human, then we really need to understand how language works and why we have it. But I suppose that kind of begs a possibly more important question, which is, why is language so special? We do know that basically all living things communicate. Bacteria communicate to each other.
00:03:54
Speaker
Flowers communicate the location of nectar to bees. So communication isn't special to
Infinite Ideas and Finite Means of Language
00:04:03
Speaker
humans. It's absolutely ubiquitous. But the kind of communication that we're able to do with language is, it seems to me, radically different from the communication that we see elsewhere in nature. And that's simply that with language,
00:04:23
Speaker
we can talk about anything. And that's really, really unique. There's no other species that we know of that can just spontaneously convey a message that's about anything at all. And if you think about it, that's really extraordinary. How on earth are we able to
00:04:46
Speaker
just reach into the kind of bottomless well of ideas and pull one out and then make a bunch of noises or move our hands and bodies about if we're using a sign language and convey that idea to another human, maybe one someone we've never met before.
Cultural Transmission's Role in Language
00:05:06
Speaker
And I've really be really pretty confident that we'll get that idea across. There's like a certain
00:05:14
Speaker
What perhaps makes it even more surprising is that of all the kind of animals on the planet, or all the living things on the planet, humans are pretty...
00:05:22
Speaker
We look pretty similar to each other, right? So we come in very set shapes, whereas plants will grow in all sorts of forms, and my dogs are way more varied. So there's something very finite, I guess, in the way that we're arranged. And yet we have this system, which is kind of infinite in its capabilities, it seems, and can be constructed in any different way. And just to give a little precursor, I think, of what's going to
00:05:51
Speaker
what's to come is that it almost seems like the finiteness, not just of us, but any kind of living thing is part of the story of what causes language to pop up, as it were, to emerge out of nothing. Absolutely.
00:06:11
Speaker
And I think, as you say, we will get there. I mean, I think when we look at that fact, we're immediately faced with this kind of duality, which I think is a really important duality between
00:06:28
Speaker
us as a biological entity in the world, you know, humans, and our behaviors, including language, but also all sorts of other behaviors that we engage in, that seem to sit somehow apart from our biology. So the languages that we speak, like the language I'm using now, English,
00:06:59
Speaker
obviously it has risen somehow from the biology of English speakers, but that seems like not a great, or not a very satisfying explanation for the facts of English, right? That, yeah, it came from a bunch of humans. So there's something else going on with language, which is quite different from, say, the signaling system in bacteria.
00:07:29
Speaker
there's another part to the story that isn't purely biological for understanding the nature of language. And that part is culture. So one of the things that really interests me about language is that it's clearly something that's
00:07:52
Speaker
part of culture in the most broad sense of how we define culture as being the set of behaviors that are passed on by social learning from one generation to the next. So language is part of that. The language I'm speaking now, I am speaking because I was exposed to it when I was growing up by my parents and the other people in my speech community.
00:08:19
Speaker
And my kids speak English because they grew up around me and so on over generations. So languages are transmitted through this thing that we would call cultural transmission. But
00:08:39
Speaker
The other thing is that culture itself, a lot of the aspects of human culture are themselves carried by language. So we have this really interesting, quite complicated, maybe we'll get into it, relationship between the cultural products of humanity, everything like the computer I'm using now to talk to you.
00:09:03
Speaker
being ultimately possible because of language, because we're able to convey instructions and able to support these very elaborate cultural behaviors. But that system itself is culturally transmitted.
Language's Power to Represent and Shape the World
00:09:22
Speaker
So we've got this strange tangled web of things going on there.
00:09:28
Speaker
So yeah, just a quick tangent. John Searle, philosopher of language, so coming at this from a very different angle, he pointed out that language has this strange, you can sort of think of language as either, as often as a mapping of the world to words. Whereas in his view, intentions are almost the opposite. When you want something, you're sort of trying to map
00:09:58
Speaker
the world to, you're trying to map your intentions to the world, you're trying to make the world like your intentions, whereas you're trying to make your words like the world. But actually, language makes things go both ways. So you can actually say something and make it true. Like one of his classic examples is like war is declared. Like that is the declaration that makes the statement true. So
00:10:24
Speaker
It's a very, very small instance. But as you say, there's this very complicated interplay between language shaping all the things that we do, and then those things shaping back on to languages that we have. Yeah. As we start to look at that kind of interaction between the world, the human world, the world we're actually living in, and language, and we see these kind of
00:10:54
Speaker
these kind of profound multi-way connections between the two, I think it becomes more and more apparent that we have to accept that an understanding of language and an understanding of humans actually that is purely biological is really not up to the task of really mapping out what it means to be human. But equally,
00:11:22
Speaker
I'm not espousing a view that says that we have to just sort of leave our biology behind and try and understand humanity purely as a cultural phenomenon.
Cultural Transmission vs. Biological Evolution
00:11:31
Speaker
I don't think that's right either. I think what's exciting about the field I work in, evolutionary linguistics, is that we're trying to figure out, we're trying to understand language
00:11:45
Speaker
And in the broader sweep, we're trying to see how it fits into our understanding of the natural world and try and understand what happens when a creature evolves that has this new system that supports this
00:12:12
Speaker
the emergence of this cultural process on the planet and how that affects the biology of humans and how human biology affects that process and what happens when that process gets up and running and what things have changed because of that. Yeah, I think this is probably a good moment to sort of lean into I guess the core of your work which is sort of making good on that promise that
00:12:43
Speaker
those wonderful capabilities of language, that structure that emerges, which gives you the compositionality and the ability to put together concepts in arbitrary ways, that comes from, or is related, yeah, comes from the process of cultural transmission. That's a very bold claim. How do you, yeah,
00:13:13
Speaker
how can we see that? We can't see that, or can we see that? Like by looking in the past and looking at, I don't know, artifacts that have been left or? Yeah, so I'll lay a little bit of groundwork first. So you mentioned compositionality there, and I think that's a good place to start. So
00:13:34
Speaker
I've said, with language, we can convey an unlimited set of messages and meanings to people. And the way we do that is actually incredibly simple. And it's so simple. It's very surprising to me, anyway, that we don't see it all over the place in nature. And the way we are able to have this kind of
00:14:02
Speaker
enormous expressivity in language is simply the fact that language is made up of parts that can be recombined. It's really as simple as that. To put it another way, language has words, right? It seems almost banal. It's so obvious and simple. So I can explain, I can talk about new things because I can take words that you and I both know
00:14:30
Speaker
and put them together in new combinations. I mean, it really is that simple. And by doing that, I can convey a new meaning that's never been uttered before in the history of the planet. And you can actually do that. You probably would do that every day without realizing it.
00:14:51
Speaker
You're probably, every day, some large proportion of what you, the noises you make will have never been made before, and the meanings you'll convey will have never been conveyed before. And it's just like, no, it's not an event. Yeah, no one's giving us like prizes for that. Yeah, you know, running downstairs and sort of made a new message, made a new sentence. So,
00:15:16
Speaker
I mean, that seems like a very simple design trick, right? Like make your communication system consist of a finite repertoire of parts and allow them to combine together and then compose meanings out of the meanings of those parts. But with a handful of kind of interesting possible exceptions, that is uniquely human.
00:15:45
Speaker
So this compositionality, which is the technical term for that, was definitely something that we wanted to try and understand the origins of. So back when I was just finishing up as a PhD student,
00:16:02
Speaker
and was getting interested in some of these ideas.
Simulations in Language Evolution
00:16:05
Speaker
I mean, the dominant way of thinking about that was to imagine that we had some kind of organ, like a biological faculty, cognitive piece of machinery that would have encoded in it
00:16:27
Speaker
the capacity to do that, and that really trying to understand the origins of something like that was about understanding how the genes that would create such an organ would proliferate in a population. So notice that that's an entirely biological take on things. You have a brain that can't do compositionality, and then some genes are favored that
00:16:57
Speaker
create a brain that can do compositionality. And once you've explained why those might be favored in this population of primates, then you've done your explanation. But that sets aside this whole thing that we've been talking about, which is the idea that language is transmitted from one generation to another. English that's compositional, right? And English isn't created by my genes, not directly.
00:17:26
Speaker
I learned English because I do have the cognitive machinery that allows me to learn English. And I use English because I have the cognitive machinery that allows me to use English. But the details of that language don't spring from my genes. They come from the fact that English is a cultural phenomenon that's passed down over generations of people.
00:17:50
Speaker
So I was interested in, well, what does this other process, this cultural process, add to the explanation? And how can we investigate it? So this was back in the 90s. And the only way I could think of investigating it was to try and build models of the process of cultural transmission.
00:18:19
Speaker
in computer simulations. So I would try and come up with simplified models of language learning that I could implement in the computer simulation. And then I would make a population in the simulation of miniature language learners and then have them communicate with each other, have them go through their lives, be born and die.
00:18:47
Speaker
pass on their language to the next generation and so on. And the idea of these simulations is you start them out with miniature languages that didn't have this compositional structure. So basically everything that needed to be said was said with a completely unique set of sounds. So there was no parts that were reused. And then just, I would set the simulations off. At the time,
00:19:17
Speaker
Like one run of the simulation would take two or three weeks. It was a very frustrating process to try and debug my code because it would run for like 10 days and then crash. I would try and figure out why it had crashed at a particular point. But to my delight, what happened again and again
00:19:46
Speaker
under all sorts of different assumptions and all sorts of different models of learning and so on, is that compositionality would emerge out of this process of transmission from one generation to the next in the computer models. So that was a real kind of goosebump moment. And then the real work began, which was to try and understand why.
00:20:15
Speaker
Yeah. I guess I actually tried this recently with chat GPT because I thought, look, I can just, I'll give it a word list and I'll, you know, assign some random set. I was doing it in hexadecimal. So I'll just give some random symbols that it needs to learn as words to one instance of chat GPT and ask it to learn them. And then, you know, the,
00:20:42
Speaker
read those back to me. If it makes any mistakes, I'll put them into the next instance. And I was hoping to reproduce your whole chain, but the problem was that it just didn't make any mistakes. So I guess the key to your, and actually I did then ask, cause you can, you know, be very specific. I said, I actually make some mistakes, but the mistakes that it made were just too regular. So it would just, it would take my symbol and it would just offset it by, it would just move it by one. So it would take like
00:21:10
Speaker
a number in a hexadecimal and move it. And it would just consistently get everything wrong by one. So it's somehow not doing what humans actually do and what I presume you must have built into your simulations, which is make some mistakes, right? But I suppose make mistakes in a particularly kind of human way. Right. This is great. We're getting to the heart of the matter. So yeah.
00:21:40
Speaker
If you think about evolution, evolution is driven in some sense, without mistakes, without error, there's no change. So if something is, if you have some evolutionary process that has perfect copying, then nothing will happen, right? So in biology,
00:22:04
Speaker
change happens partly due to mutations where a cosmic ray will hit a piece of DNA and change something about it, or by the fact that the inbuilt into sexual reproduction is recombination. So we end up with kids that aren't, thank God, clones of their parents.
00:22:29
Speaker
So biology has this process that generates variation, generates change. And of course culture does too, right? Like the languages that we speak are slightly different from the languages that we heard and our language continues to move.
00:22:51
Speaker
So an evolutionary process has to have, we can think of those as errors, it doesn't really matter, but it has to have some place where innovation happens. And it's my feeling, and maybe we'll get into this, that the way that happens in culture and biology are actually
00:23:13
Speaker
quite different and importantly different. But some of my colleagues disagree with me. And so there's actually a very close analogy between biological evolution and cultural evolution. I don't think that's quite right. But I'll try and we'll take it right back to that first simulation I did. We'll try and work out. I'll try and explain why we got the results that we did, why compositionality emerges.
00:23:41
Speaker
And it all boils down to a decision I had to make when setting up simulations, which is basically how long is the lifetime of my simulated people? I could make them live a really long time, so long that by the end of their lives, they will have had all the things
00:24:08
Speaker
that in the little world that I'd given them to talk about, they would have heard everything expressed. There'd be no new ideas for them to observe. If you do that, then basically you end up with very little happening. So my simulated agents actually had perfect memory. And if they heard something in their world being talked about by another,
00:24:38
Speaker
simulated individual, they would remember how that was expressed and just store it. And so when it was their turn to talk about that thing, they would just reproduce it. And so there we have this kind of perfect copying, right? So this initial random unstructured language we gave the agents, which is, agents is what we call a simulated individuals. We give these agents this random unstructured language at the start. If their life, if their lives are too long,
00:25:09
Speaker
then they just learn it perfectly. And over generations, nothing happens. So there's no evolution. So what I did was I shortened the length of the lifespan of my simulated individuals so that they weren't guaranteed to hear every possible sentence. Right. So
00:25:33
Speaker
So what that would mean is sometimes they were called upon to say something that they'd never heard said themselves, which if you think about it, that's exactly what we do in language all the time. So the space of possible things that we might want to talk about is vastly bigger.
00:25:52
Speaker
than the set of sentences that we've heard in our lifetimes up to this point. What that means is we're constantly generalizing. We're constantly trying to go beyond the data that we saw to talk about new things. And so I set it up with this parameter, which is basically the lifetime of the agents, so that they too had to talk about things they hadn't had talked about.
00:26:20
Speaker
And so then when you think about what happens then, so you're suddenly asked to say something that you don't know how to say. And what my agents did is if they had no idea how to say it, they would just produce something, they would just produce something random. So I built in this kind of like random invention process. So that produced this,
00:26:48
Speaker
process of change over generations in the simulation. Now, if you think about that, what's happening, right? So you've got a random language it starts off with. It's got no structure. There's no parts that are being reused. And then it changes because the agents don't see all of that language. And so they have to generate new random sentences.
00:27:12
Speaker
And so the next generation here is another random language. So it feels like nothing's going to happen at interesting here. And when I set this simulation up, I kind of thought, OK, this isn't going to work, but I'll try. It feels like nothing should happen that's interesting. It should just like endlessly be lots of random utterances. But the lovely thing that happens is that sometimes,
00:27:41
Speaker
Just by chance, you could have two sentences produced by one of the agents that were, let's say, about something similar in the world, in their little simulated world. And the two sentences might share, just by chance, some aspect of their signal. So there might be some little bit of what they said that was similar between these two sentences, but just by chance, corresponded to something in the world that was similar.
00:28:10
Speaker
And what they would do sometimes is that they would make a mistake and they would think that this little piece of signal actually corresponded to that little thing in the world. So they would kind of hallucinate, if you like, a word where there hadn't been one before. It was just random chance, but they think it's a word. And what happens then is that every time that agent wants to talk about that thing, it uses that word.
00:28:40
Speaker
So it reinforces its... Exactly, exactly. And so then the next generation is exposed to a language which suddenly has this repeating little subunit in it because it was generated by this individual in the last generation. And so they learn that subunit and it persists over time. And so what happens is very rapidly,
00:29:09
Speaker
you start getting these pieces of structure and once they're there, they stick because they're easier to learn. But it's all driven by this kind of finiteness of the lifetime of the individuals. Because if there wasn't that, there wouldn't be this pressure to
00:29:29
Speaker
to come up with ways of communicating that can survive through this limited lifespan of the individuals. Yeah. And some people might be thinking, oh, well, congratulations. You've proved that computers can generate language, right? But, you know,
00:29:55
Speaker
I've made enough coding errors to know that sometimes the things that I've written when I try to actually look at the real world, my simulation doesn't quite match it. But remarkably, you've kind of done similar experiments with real people. Maybe take us through one of those. Yeah, so I guess that there was a few years after that kind of those simulations where
00:30:24
Speaker
I was immediately faced with that. I would go to conferences and describe this and people would be like, well, that's nice. If they understood at all.
00:30:37
Speaker
And so I would try and I would first first approach to that, that problem was I was trying to understand the generality behind it. So like, what is this general process that creates this structure? And so I would do this thing with lots of different, I would rewrite the simulation from scratch using completely different models of learning. So using neural networks or use something else or whatever. And I would try a bunch of different ways of doing it and then
00:31:04
Speaker
other people in different labs tried to replicate the result and so on. But it was slightly frustrating because it did feel, you know, we ended up having like conferences where there would be a room where the computer modelers talk to each other. Right. Literally, the evolution of language conferences, there would be a track that was for the computer simulation people.
00:31:34
Speaker
And so we had a great time kind of, but it also felt like we weren't really having the impact we felt we ought to have because we were just talking to each other. And it didn't seem that people really understood what the hell we were doing. And one of the problems is that it became more and more technical and less and less clear to your average.
00:32:00
Speaker
like linguist or evolutionary biologist, what on earth we were playing at.
From Simulations to Human Experiments
00:32:08
Speaker
But to me, the insight that I really wanted to hold onto was this very, very simple idea that there are structures in language that
00:32:24
Speaker
are it's possible to learn with a finite set of sentences given to you. And because there's a finite set of sentences given to each generation in cultural transmission, those properties of language are actually inevitable. So there's actually an insight into cultural evolution that says wherever you've got a learner that has to generate
00:32:52
Speaker
generalized from finite subsample of data, generalizations will emerge over generations. And that seemed to me to be an incredibly general point that isn't just even just about language. But we still weren't getting traction with all of the simulation work. So we took a term, I guess,
00:33:23
Speaker
about 10 years later, that we moved from doing work in computer simulation to doing work in people, real people. And the story behind why that happened is quite a funny one. So I'll give you a little anecdote of how we ended up doing experiments with people instead of simulations.
00:33:50
Speaker
I had a quite brilliant MSc student at the time, Hannah Cornish, and she was doing an MSc in developmental linguistics in Edinburgh.
00:34:06
Speaker
And she'd done with me as an undergraduate student some computer simulations of language evolution. And that's what she wanted to do her dissertation on. So she wanted to do some of these cultural transmission experiments, which we call iterated learning experiments, by the way. In simulation, she wanted to do this. But her program director said at the time that MSC had a rule that said you had to do your dissertation
00:34:35
Speaker
with an experiment on human participants. But she wasn't allowed to do her planned study in simulation. And so she came to me and said, I'm really sorry, I'm not going to be able to do this project because I have to do an experiment. But she is quite...
00:35:02
Speaker
stubborn, and we both are, I think. And we thought, well, let's just do exactly what we planned, but do it with people instead of these pieces of computer code. So we set up exactly the same design, but where she was planning to have, like, I think neural networks
00:35:30
Speaker
we would just get people into the lab and teach them these kind of unstructured languages and then get them to produce sentences from the unstructured language that we trained them on, and then take the output of each of those participants in the lab and give it to the next generation of participants who came into the lab. So we just did a drop in, like,
00:35:59
Speaker
replace the internet work with people. And that turned out to, that MSC dissertation turned out to be my most cited paper, so. Yeah. Because what we found was exactly the same phenomenon. We get these unstructured languages
00:36:24
Speaker
becoming compositional over generations just through this process of transmission from one generation to the next. Yeah. One of the experiments that I really like, because I think it makes it very clear exactly how people can latch on and create a structure which allows them to even figure out words that they've not seen before.
00:36:54
Speaker
And I think the experiment I'm thinking of, and I don't know if it's the one you're referring to, Hannah Cornish, but it's the sign language one where you have a few kind of different themes. So you've got, is it religion, food, and you've got this kind of two dimensional structure of categories. So you've got themes down one axis, and then you've got functions down another. So one function might be,
00:37:20
Speaker
place or one functional category might be place. So the place for religion would be a church, maybe another functional category was maybe action. So that could be to pray or to preach. I can't remember exactly the ones that you picked out. But you had this kind of matrix of and then again, you start with every you teach people a sign language where everything there's there's 16 combinations because there's four by four.
00:37:51
Speaker
Grid and each one of those 16 words is like completely different sign. Yeah. And first generation is taught it. They have to reproduce it and the symbols that they reproduce are given to the next generation and so on. And yeah, you should describe the results because. Yeah. I am very fond of that experiment. I think it's but that was work by Yasmin Motamedi. I think it's a really beautiful study.
00:38:22
Speaker
And the motivation to move into that gestural modality was partly because the other place where you can see this process I've been talking about happening, not in computer simulation and not in the lab, but in the real world, is the cases where new sign languages spontaneously emerge. And there's about, I think, 60 or 70 of these right now
00:38:52
Speaker
currently that people have discovered these spontaneous sign languages and made these occur in the world due to either deaf kids of hearing parents who are brought together in a school context for the first time or because
00:39:14
Speaker
because of a gene for deafness spreading often in a very remote population and suddenly a large number of deaf individuals being born in that village, say. And what's really remarkable is that we go very rapidly from something that looks kind of arguably less structured than
00:39:38
Speaker
a typical human sign language, typical more mature sign language into something that looks very, very familiar to anyone who has studied sign languages.
Sign Language and Cultural Transmission
00:39:50
Speaker
So that work had originally been, those results had originally been kind of
00:40:00
Speaker
hailed by people who wanted to say that the capacity for language is innate. It's part of this biological faculty that I was talking about earlier. Even when the culture, there isn't this culture of shared language, we still spontaneously create it. But what I feel quite strongly is that actually it's a demonstration of the power of this cultural process
00:40:29
Speaker
It's very rapid, but we can actually observe it happening in these situations. So we wanted to recreate that process in the lab as well and have participants do this iterated learning kind of experiment with this gestural modality. Ours is a lot of fun. Basically getting people to play charades for a living is great.
00:40:56
Speaker
The thing that we discovered with that study that I think is really important and actually added something to what we'd found 20 years earlier with the simulations was that there are two really important processes at play that create this structure that we've been talking about, this compositional structure.
00:41:22
Speaker
Previously, we'd been thinking really that it's learning that matters. It's the fact that you have to learn the language in each generation. But that study and a couple of others that we did around about the same time showed that that wasn't the only pressure that mattered. There's another pressure which is coming from communication.
00:41:45
Speaker
So what Yasmin did in that study was she compared a situation like you described where people see a bunch of gestures for a particular meaning and then they have to reproduce those gestures and then the next generation sees their gestures, video recordings of their gestures and has to reproduce them.
00:42:09
Speaker
So that's what we would have thought of at the time as a standard iterated learning process, where you see this set of gestures that's been culturally transmitted from one generation to the next by learning. Now, what we find with that setup is actually quite weird. The gestures start to look really over-elaborate. So over time,
00:42:39
Speaker
they became longer and longer, and they had arguably, they had parts. So sometimes you would see gestures. You talked about church, for example. So a gesture for church might have a part that corresponded to building and another part that corresponded to religion, but they were really unsystematic and they were all over the place and they were very long. So, and each generation,
00:43:09
Speaker
made the set of gestures longer and longer and longer, more and more kind of baroque, didn't really look particularly structured. It didn't look like the kinds of things that we were expecting, where we would have expected this, like you talked about this matrix of meanings. We were expecting really two-part signs, one for the kind of religion and another one for building, let's say.
00:43:38
Speaker
So the other thing that Yasmin did was she got people to do this experiment and instead of transmitting it down over many generations, she got pairs of participants in and they had to play a little communication game with each other.
00:43:54
Speaker
you know, much more like a game of charades. So one participant would be given a meaning to convey, and then the other participant would watch and have to guess what the meaning was. And then they would swap roles. And they were both shown this initial unstructured set of gestures at the start.
00:44:13
Speaker
And what happened there also was weird, very different, but also didn't look like a sign language. So what happened there was that the gestures got really, really short and really efficient, and they would be completely arbitrary-seeming. So one gesture would just be a movement of the hand, and then it'd be church. So you've got really good at understanding each other.
00:44:41
Speaker
Yeah, when you look at it, what's going on there is that these two participants having a lot of time playing together and they get this shorthand for conveying all of the meanings perfectly and they've left.
00:44:58
Speaker
So neither just the process of transmission by learning, iterated learning, nor the process of communication between the pair for a long time was sufficient to give the kind of structure we were looking for. But if you put both of them together, then it works. So the kind of real experiments
00:45:23
Speaker
experiment that gives the result that's exciting is where you have pairs of participants playing this kind of communication game with each other and then a new pair of participants come in and see the gestures of the previous pair and then play the communication game and then a new pair of participants come in and see the gestures of the previous one. So you have this kind of
00:45:49
Speaker
process whereby there's communication within a generation. And then there's cultural transmission across generations.
Cultural Evolution and Optimal Language Design
00:45:56
Speaker
And then you get this like, absolutely crystal clear system where there's one sign for like religion, another sign for building. And all of the signs, three parts, very crisp, very efficient, very structured.
00:46:19
Speaker
Yeah, so effectively, you just need kind of eight symbols in this example, like one, four for the theme, religion, food, et cetera. I think one was photography. Interesting, exactly. And then one for the kind of function. Well, actually, the cute thing, that's what I would have said too, but actually the cute thing is you only need seven. Okay, so that's one, that one, yeah.
00:46:48
Speaker
Yeah, so in fact, that's what happens. So the system that comes out, it has one of them as a default. I guess that's what we call one hot encoder. Yeah, right. And it's really striking to me that these cultural processes can come up with optimal design
00:47:13
Speaker
really efficiently. And it's like no one is inventing this stuff, right? There's no one in there going, I know, like,
00:47:22
Speaker
if we have to describe these 16 meetings. Because they're just there for 15 minutes. And they're not going like, here's how we should design it. We should take one of these and make it default, and then not have a second part for that. And then all the others should have it. There's no genius in there creating this system.
00:47:44
Speaker
But the collective actions of this population in the lab is creating a linguistic system that is like exquisitely well designed for the two pressures that we've put on it, which is to be useful for communication and to be learnable by the next generation. And again, this feels to me like a kind of important lesson about culture is
00:48:15
Speaker
that culture can optimize and jointly optimize for these various pressures that we put on it. And these optimizations are not the product of design. They're not the product of intelligent design on behalf of the carriers of culture, us. And we've seen that again and again, actually. So we have a whole host of experiments where we change things about frequency
00:48:46
Speaker
We change the signal channels so that it's costly in various ways. And the languages that come out are just these beautifully optimized systems. And I think that really made us think a lot about where the engine for
00:49:12
Speaker
explanation of human behavior should lie, you know, and, and yeah, I guess, I guess sort of teaching us that culture is this incredibly powerful computational system in its own right. And I'm not sure if I totally understand yet the implications of that.
Human Limitations and Language Evolution
00:49:32
Speaker
They go beyond language, though. Yeah. Yeah, I think I want to lead back to
00:49:39
Speaker
the point I made very early, which is just that it's the generational aspect, but also the fact that within each generation, there's a kind of finite amount of time that people spend collaborating, communicating with one another. And also at the kind of generational shifts as the fallibility of our learning. And so in both cases, it's, you know,
00:50:09
Speaker
because we are quite limited beings, that language evolves. And one of the surprises here, and we can talk about this presently, is those limitations we seem to share with our other animals. So there's a little bit of a mystery here. As you say, we get this wonderful system, but it almost seems too easy.
00:50:37
Speaker
I think one thing that one might think here is, oh, this is to do with categories. So I think in the philosophy of language field, people get really hung up not so much on the compositionality and those things about language, although it's of lots of interest, but they're just really interested in, I guess, how language is managing to represent stuff and that we have this
00:51:07
Speaker
Yeah, we have this model of the world, which is language. And one might say, well, the reason this experiment works is that humans already have that kind of categorization into function and into theme. And so it makes it easier for them when something kind of maps that. And with your kind of computer simulations as well, I guess you said, well,
00:51:37
Speaker
when the program or the agent was presented with two kind of stimuli, two things that it was trying to express that were similar to kind of states of the world, which are similar, and it happened to produce linguistic expressions, which were similar, that would be striking to it. But obviously, you know, you have to, and there's no way around this, you have to kind of sneakily encode, well, not sneakily, but you have to encode into
00:52:04
Speaker
agent that it has this kind of concept of similarity in states of the world. And I don't know if there is any way of explaining that. Or if that's uniquely human either one expects that animals can also have like, you know, they seem to have pretty good concept of time and space and probably and of similarity and, and therefore that doesn't seem particularly unique. But it is. I mean,
00:52:34
Speaker
That might be just one of those hard problems that can never be sort of like how. I mean, I think you're thinking along the right line. So one way of thinking about how to tackle these kind of questions is to think, what are the minimal requirements to get something like this process that I've talked about up and running? And that's that tends to be where I think the simulations are really good. Like, like, as you can say,
00:53:02
Speaker
How dumb can I make my simulation and still see this kind of thing happening? And so for compositionality, I think the sorts of things you need, you need some measure of similarity. You hit the nail on the head there. You need a notion of similarity between things that you're expressing, the meanings of things out there in the world. And you need a similarity measure over signals.
00:53:32
Speaker
So you have to have two signals. All the signals can't be completely distinct from one another because then there's nothing for the structure to hang off. So those are two things you need, two basic requirements. I think they're both very achievable by pretty much anything. It seems inconceivable to me that
00:53:57
Speaker
any animal could be successful in the world if it didn't have some similarity metric for the world. I mean, just really bizarre. Whether or not they, I mean, for signals, that's a little bit more of an open question, but you would need that too. Then the other thing you need is you need a bias
00:54:25
Speaker
for representing a mapping between those two spaces, the space of meanings, the space of signals. And that bias has to prefer simpler mappings. It turns out that that's a key thing that drives all of this, is that the simplest mapping between a world of meanings that has some similarity metric on it and a world of signals that has some similarity metric on it, the simplest mapping between those two
00:54:55
Speaker
is one preserves similarity from one space in as much as it can in the other. So a compositional language is one which is similarity preserving. So similar meanings have similar signals. I guess it's quite easy to conjecture that, and I suppose you don't even need to conjecture because you can see it in your simulations, but I would guess that the pressure to have a simple mapping is almost like
00:55:25
Speaker
the same as the pressure to make things easily learnable, like to have. Exactly. Exactly. Exactly. So, so, so the obvious question that sort of leaps out of all of this is like, how much can I assume these three things as being just sort of like, I don't have to argue for them, I don't have to mount some special pleading that humans have these sort of
00:55:54
Speaker
And I actually think any animal would have these. So there is a line of thought that says that basically all learners have a bias for simplicity. So there's work by people like Nick Chater and colleagues that talk about a kind of universal bias for simplicity. And you can see it in things like Occam's razor, which is the idea that we should prefer simple explanations.
00:56:22
Speaker
So there's a sense in which this is sort of things that we're allowed to take for granted is that it's reasonable to assume that learners will be biased towards simple explanations. And a simple explanation of a mapping between two structured spaces is all that we need in order to get this process to work. Now, it doesn't give you structured language. You still need the cultural process.
00:56:51
Speaker
which has to include learning and communication in order to get structured language. But if you've got those elements in play, then that'll work. Now, so all that kind of argumentation kind of
00:57:11
Speaker
feels like it's pointing towards the fact that structured language should be absolutely ubiquitous. It should be everywhere. We started with saying this is uniquely human and it's our special trick.
00:57:28
Speaker
And so it feels almost like I've massively shot myself in the foot, right? I've mounted an argument. It's like they should just be everywhere. So the very thing that we were trying to understand suddenly becomes all the more mysterious. But I think the way to really address that question or to get at that puzzle
00:57:57
Speaker
is to think that what we've actually done is shifted the goalposts. I don't know if that's the right metaphor, but we'll go with it. So previously, the question was, how did our species end up getting this compositional structure, let's say? How did we end up with a biology that somehow encodes the structural properties of language?
00:58:26
Speaker
And instead now I'm saying, how did our species end up with this engine, this cultural engine that creates language structure? And actually that's completely different. And I think a much more tractable question. So in order to get structured behavior off the ground, what we're saying is we need cultural transmission and
00:58:55
Speaker
of a particular kind. We need cultural transmission that has this iterated learning aspect and this communicative aspect. So now we can go and look at other species and say, well, where do we find these things? Where do we find learned signaling systems?
Cumulative Culture and Learning Biases
00:59:13
Speaker
Are they transmitted culturally? Do they involve communication in the sense of communicating about meanings and so on and start looking for
00:59:25
Speaker
for these kind of pieces of this cultural machinery elsewhere in nature. And that's where I think we start to see what's really needed, what biological evolution needed to deliver for us as a species in order to get this cultural process off the ground. Yeah, so we need to be learners.
00:59:57
Speaker
I've sort of, I recently came across this one of experiments you're probably aware of. This might seem like a tangent, but I think it does address the point. From Victoria Horner and Andy Whiton, just based over the first in St. Andrews, where I think it's maybe a couple of decades old, but they looked at chimps and they had this experiment with
01:00:25
Speaker
a box where you showed a chimp that if you poked in the top of the box and the side of the box, a little treat would come out. And they showed this to chimps and children. And they both learned to poke a hole in the top and poke the stick through the side. Poke the stick through the hole in the top and then through the side and you got a treat. But then they repeated it and they made it a transparent box. And it became apparent that
01:00:55
Speaker
You didn't need to poke the stick through the top hole. It's just poking it through the side hole, which produced the tree. And the chimps, clever that they are, they're like, OK, well, I'm not going to bother poking the stick through the top hole. I can see that that's doing nothing. But the children still did both. So in going against all of the empirical evidence, these dumb kids put a priority
01:01:25
Speaker
I think it's called over imitation on copying what another human was doing. And that, you know, maybe that's one of the kind of uniquely human things that then sets the chain in motion, I guess, to this, the development of language. Yeah, I mean, I think that's, I love that work. And that work is often talked about in the
01:01:52
Speaker
context of cultural evolution or generally saying that although culture isn't uniquely human, there are a lot of species that have it. It's not massively common in other species, but it is there, which is great because that means we can actually start doing some comparative work looking at culture in different species and see how it's similar on different human culture.
01:02:23
Speaker
But the culture of the evolution folks emphasize cumulative culture as something that seems either uniquely human or at least rare elsewhere. Cumulative culture is the idea that we build cultural products that over multiple generations accumulate often in complexity. And so without cumulative culture, we wouldn't have laptops and things, right?
01:02:50
Speaker
it has to build on multiple kind of innovations that have gone before. So one of their questions is like, what's needed for cumulative culture? And one of the things that has been argued in that literature is that it needs very high fidelity copying. And the over imitation stuff has been used as
01:03:14
Speaker
as possible evidence for very high fidelity copying as a kind of bias that human infants have and humans generally actually have the desire to copy. I think there's maybe a little bit more thought about with that, but I do think it's very relevant to language.
01:03:40
Speaker
Actually, what I think we need to be doing now is really thinking about precisely what kind of learning is needed for language to take off in the way that I've described. We actually have to get into the details. And it is that learning, is that type of learning, the sort of learning that we see in these over imitation experiments. So there are people working on
01:04:09
Speaker
really getting into the details of what's going on in language learning and how that relates to learning another species. So people like Inbal Arnon, who is working on, who's a lot of work on the idea that
01:04:28
Speaker
language learners, infants, have this bias to start learning holes and then only later discover parts in those holes, which actually aligns really well with the results of a lot of our experiments and our simulations. So one of the things I think I suspect that's coming in the next few years will be a kind of mapping between what we know about language learning
01:04:57
Speaker
onto social learning more broadly in humans, but also social learning in other species. So I think this is very exciting because we're actually starting to get into the nitty gritty, exactly what kind of cultural transmissions needed for language. And exactly those kinds of experiments, over imitation experiments, give us a kind of clues as to what's different across species. And
01:05:27
Speaker
Yeah, so then we can do the next step, which is to say what biological mechanisms support that kind of learning? And how might they have evolved? What are the evolutionary pressures that gave humans this propensity for a particular kind of copying? It might have nothing to do with language. I mean, I personally think it's a reasonable hypothesis that language isn't
01:05:57
Speaker
accidental byproduct of changes in the way in which we learn information from conspecifics. It's not out of the question that none of the adaptations are there because of the consequences they had for language.
Language's Non-Utilitarian Roles
01:06:25
Speaker
I mean, it's certainly very hard to imagine language evolving as nature's trying to get us better at expressing some signs that map onto the world. So it's kind of like a spandrel, which has come about because we've just got very good at passing, copying useful behaviors from one another. And we get so, I suppose, trusting at that, that even when the behavior doesn't seem useful, when we can't see the utility, we'll do it.
01:06:56
Speaker
I mean, that maps very well onto the fact that almost all of our knowledge comes from testimony and language rather than from things that we see in the world, isn't it? I mean, I actually think there's some beautiful experiments that suggest that language has got a lot less to do with useful information transfer than we like to think. There's some lovely experiments by Gareth Roberts and colleagues where
01:07:27
Speaker
They had pairs of participants communicating over a chat interface for, I think, an hour or something. And they're just typing away, just chatting about whatever they wanted. And unbeknownst to the participants, halfway through, they swapped partners. But this was completely... So you would suddenly be talking to somebody who didn't have a half-hour conversation about something else.
01:07:57
Speaker
This sounds like a brilliant, it's just conspiracy theory or something. I just wonder if you've been somehow replaced mid-conversation now. The funny finding was, you know, obviously quite a few people noticed, but not everybody. And there's a sizable proportion of people when asked, did anything odd happen in that conversation? They go, no.
01:08:23
Speaker
I think that tells you that a lot of what we're doing with language isn't utilitarian. Yeah, that's right. I think it's very playful. The next recording that I'll do for this podcast is going to be with Gordon Burghart. He's sort of an expert on play. One of his major contributions was
01:08:53
Speaker
animal, animal play in particular, I should say. Although he, I mean, he's always at pains to point out that humans are animals too. But one of his big contributions was trying to come up with a definition of play that actually is useful because we tend to just, you know, we think we can recognize play. And, and yet
01:09:18
Speaker
For many years, there were particular species that we didn't think could play. And we only had a very loose, even though we could recognize it, if you can't define it, it makes it hard to build academic fields around it. So he came up with this five criteria. And one is that the activity that's being done is not fully functional in the context that it's expressed. And you can really see that with where I'm going with this. Language learning looks a lot like play.
01:09:49
Speaker
in that it seems to me many of his definitions. So when children just randomly repeat the words that you're saying, particularly when they're very young, they're not doing it because they're just imitating, right? And play is often very imitative. Another of the criteria is that it has to be a kind of repeated activity, but with changes made in the repetition.
01:10:16
Speaker
so that again, children will say things over and over and like kind of mess around with the words and we'll actually do that ourselves as well, even as adults. And another is that it's sort of precocious, like it prefigures something that you need later. So there's just so many, you know, to me, a lot of language learning and even language usage seems to fit this
01:10:44
Speaker
Yeah, to fit play. Certainly when we're talking, but we're not passing around important information, which is, let's face it, most of the time. It's hard to see the functional adaptation, right? But yeah, it's producing something for us. I think, you know, Bocat's ideas here are that, oh, yeah, finally, another important thing I wanted to mention is that, yeah,
01:11:10
Speaker
play happens under conditions free from stress and sort of disease and other evolutionary pressures. And I think that's kind of part of this key to unlocking where this over imitation can happen is that we seem to have this kind of excess of resources, which other species don't have. And that kind of relaxes a lot of the things that would be impeding our play, impeding our learning.
Baboons and Cultural Evolution
01:11:38
Speaker
Yeah, I think that's that's a good insight. And I like that direction of thinking. So just maybe something that supports that view. So and along the lines of like, what are the minimal requirements that we need in order to get something like this kind of systematic structure and behavior emerging?
01:12:03
Speaker
So we've actually collaborated, so myself, Kenny Smith, and Nicolas Cledier, we collaborated on an attempt to recreate this kind of iterated learning experiment in baboons. So baboons are not a species that has this kind of behavior in the wild.
01:12:32
Speaker
They don't have this systematically structured behavior like we see in language. So we decided to try and run an iterated learning experiment on baboons, and Nicola has
01:12:50
Speaker
this incredible facility down near Marseille, where he has a population of baboons. They're obviously captive, but they've kind of free roaming and they have an enclosure. And in the enclosure, there are these portacabins.
01:13:08
Speaker
that they can come in and out of whenever they want and inside the porta cavern is a touch screen and a robotic food dispenser and the baboon can reach through to the touch screen and the computer recognises the baboon and serves them up a quick experiment on the touch screen and they get reward depending on however you design the experiment.
01:13:32
Speaker
So Niko is able to set up these experiments, like quite elaborate and complex experiments, and just like them running on this population of baboons, and get tens of thousands of data points in a very short amount of time. So we're able to do experiments with those animals that you would never do in the traditional context where you have to do experiments. This is so amazing, by the way. It's like baboon mechanical tuck. Exactly, exactly.
01:13:59
Speaker
So we ran this experiment. And we had to use a behavior that they already knew how to do. And it was a simple task where the baboon was shown a grid of buttons, 4 by 4 grid of buttons.
01:14:23
Speaker
And four of them would be lit up at random, just for a fraction of a second. And then the baboon had to press the buttons that had been lit up. And if they got most of them right, then they would get a food reward. And that was it. And they just did this over and over and over again.
01:14:41
Speaker
But what we were doing is we were recording the baboon's responses, whether or not they were right or wrong, and then using that pattern of responses and transmitting that to the next baboon in the experiment. So you start off with these kind of random light patterns and then these light patterns got transmitted as a set.
01:15:02
Speaker
It was a set of 50, I think, of them were transmitted to the next baboon who would then try and recreate them. Now, the baboons didn't know they were involved in a cultural evolution experiment. They were just pressing buttons, getting monkey chow. But they were. They were part of a cultural evolution experiment. And what happened was that the light patterns evolved over generations.
01:15:30
Speaker
And they evolved to have systematic structure. And completely bizarrely, they evolved to be shaped like the Tetris pieces that you see in the game Tetris, which is kind of wild. So you have these Tetris shapes.
01:15:56
Speaker
Now, you might think about that. Well, maybe those tetra shapes are easier for the baboon to press. And so the culture was evolving to be easier for the baboon to recreate. But what was really kind of mind blowing is that actually the opposite is true. It turns out that, and we know this because we can test the baboons on a vast number of shapes. They actually find the tetra shapes harder to copy individually.
01:16:24
Speaker
But when they appear in a set, the baboons copy them more accurately. So what's happening is that the set of patterns is evolving systematic structure so that over the whole set, they're easier for the baboon to copy. So this is another example of that kind of process of, you know, this cultural process being, it kind of has this incredible computational power that goes beyond
01:16:53
Speaker
what's going on in the individual animal's head. But that's not the reason I raised this. What this shows us is that you can take an animal that doesn't engage in this kind of cultural process in the wild, this cultural transmission of sets of complex behaviours that doesn't happen.
01:17:15
Speaker
and nevertheless get it to do something that looks actually quite similar to, at least formally, computationally, looks quite similar to what we see in language with emergence of systematic structure. So you could say, well, what did we change? What did we do to that animal that made it possible for them to do that? And what we did was we just gave them motivation to copy. We gave them the robotic food dispenser.
01:17:42
Speaker
So the animals were rewarded for copying these patterns. And so to me, that says if we were to look for a minimal requirement that humans have, is we've got an endogenous reward system for copying. We just love to copy. We just get those happy hormones every time we copy.
01:18:11
Speaker
That's something that biology can give us, has given us. It's given us some kind of wiring up of the reward system for copying. And that's unlocked this
01:18:26
Speaker
new computational evolutionary system of culture that just rolls out as a consequence of that. And we can see this elsewhere in nature. So not everywhere, but in limited, sorry, not limited, I shouldn't say that, in narrow contexts. So we can just look at birdsong, for example. So birds copy complex.
01:18:55
Speaker
behaviors in other members of their species in their song. And that is very, that's underpinned by the reward system in that species, that the birds are rewarded endogenously for copying song. And again, song is another one of these culturally evolving, complex systems that we see in nature.
01:19:23
Speaker
Yeah, I'm really glad you brought up the example of the boons because, well, firstly, I mean, it does speak to this point really nicely. And it reminds me of one of the other of the definitions or criteria for play was that it needs to be pleasurable and spontaneous. And so that's what we humans seem to get from imitating others. But the baboon, we had to kind of wire that in by giving them a fruit reward, monkey chow.
01:19:54
Speaker
And also, yeah, it's just a wonderful experiment with baboons. And the fact that, well, I learned from watching some of your talks, which I really recommend to listeners because you get to see videos of people doing sign language and all sorts of things.
Art, Music, and Systematic Structure
01:20:07
Speaker
But yeah.
01:20:09
Speaker
the set of those shapes for Tetris are called the Tetronimos and that the baboon has sort of created a grammar that matches that. And it also shows that this kind of structure, it doesn't need to, it emerges just, it's not necessarily to do with the, you can separate out the representative angle of language and just look at how,
01:20:38
Speaker
structure evolves in any kind of learning task. And that also moves us beyond language to other human behaviors like art and music, I think, which are also culturally transmitted behaviors that have systematic structure, but they don't have the pressure to communicate, at least not in the same way that language does. So they have
01:21:05
Speaker
the solutions that cultural evolution finds for music, let's say, are different from the solutions that cultural evolution finds for language because the pressures to be expressive are different in these different domains. Yeah, there's no clear categories that we're mapping out with music.
01:21:32
Speaker
Yeah, and this is a good place to end up because you're also an artist as well as a linguist. So perhaps you can tell us a bit about your work there.
Simon Kirby's Artistic Ventures
01:21:43
Speaker
I have a couple of your pieces not behind me, but yeah, you do a lot of different things, even within the art world. So tell us a little bit about that. Thank you for bringing that up and thank you for
01:21:59
Speaker
Having a couple of my pieces, I'm very flattered. Yeah, this kind of started off, again, I guess about the same time I started doing experiments on cultural evolution of language. I was...
01:22:17
Speaker
working or hanging out with some artists and designers who were doing installation work. And initially, I was just thinking that it would be fun to kind of help out on the technical side of some of the projects that they were working with. But one of the things it taught me very quickly is that the kinds of questions that artists grapple with
01:22:47
Speaker
had the same as the questions that scientists or the scientists in my area grappling with, questions about what it means to be human and what are the effect of the cultural processes on human behavior and so on. These are all sorts of questions that did not seem unusual or alien to be talking about with my artist friends. And I realized that I'd been brought up with this
01:23:16
Speaker
or sort of internalize this kind of, this idea that the arts and sciences were somehow radically different from each other and involved, you know, even we're engaged with by very different kinds of people. And I think this is incredibly harmful and just plain wrong. So I decided like I've got no arts background at all myself.
01:23:42
Speaker
because I grew up with scientist parents and always wanting to be a scientist and went through all of that.
01:23:54
Speaker
stopped all of the parts of school education that weren't servicing that goal. And I thought, well, why don't I just see what it's like to try and do this stuff? Try and tackle some of the questions that I'm tackling scientifically in an artistic world. And I was so lucky to have very, very patient and kind
01:24:22
Speaker
artistic collaborators who are willing to sort of go along with me on that ride. And it ended up with us actually collaborating on a whole series over over 10, 15, actually more, probably more like 20 years now, a whole series of projects, a lot of them that ultimately are about human
01:24:49
Speaker
the ways in which humans are changing as a cultural species, particularly recently with the emergence of things like social media. And that kind of deep, long-lasting collaboration really gave me confidence in saying it is actually useful
01:25:10
Speaker
to try and engage in artistic production as well as scientific production. And my secret mission is to, at some point, not be able to tell which I'm doing. That's my goal, is not be able to tell on a particular project whether I'm doing art or doing science. And so one of the things that I've done to try and get there is on scientific grants. In all the scientific grants, I put in an artist in residence as part of the proposal.
01:25:41
Speaker
And then the idea is not to have this artist in residence, just sort of interpret what we're doing as some kind of product. Or just paint you as you go about. Exactly. But to have them have them involved in not all of the work, but have them involved in thinking about design of experiments, actually run some experiments and
01:26:06
Speaker
So, for example, in my last grant that was looking at cultural evolution of these
01:26:15
Speaker
gestural languages, gestural systems. I was working with Tommy Perman who's a designer and we came up with an experiment that was about drawing and representing the world through iconic drawings and we actually ran that on a huge number of participants where they had to copy drawings
01:26:39
Speaker
And so the hope is eventually we're going to have an exhibit showing cultural evolution of drawings that were the product of Tommy's work on that project. And I think that it's something that would encourage everyone to think about the connections between the
01:27:08
Speaker
the kind of debates and discussions that are coming out of work in the arts and work in sciences and realize that there are all these insights that you can get from this slightly different mode of working that are relevant for the other mode.
Artistic Approaches in Scientific Research
01:27:32
Speaker
So I try and spend a good chunk of my time in artistic production. And I find it such a useful tool for resetting thinking. I mean, one of the things that happens when you're thinking in a kind of scientific mode is that there's this
01:27:57
Speaker
there's this kind of quite restrictive frame that you have to put around everything. Like you might, when you're designing an experiment, you're necessarily thinking about the analysis and you're thinking about pre-registration of the design and you're thinking about how you'll get this published and the critiques you might have and so on. And all of that, I'm not, all of that's super necessary. Like you, I'm not espousing throwing all that out.
01:28:27
Speaker
But if that's all we do, I think there's a tendency that we can box ourselves in to being conservative. There's a danger that we don't come up with creative solutions to the questions we're asking because we're so trained to be risk averse. And in an artistic frame of mind, often there's far less
01:28:53
Speaker
of those kinds of constraints. And there's a different set of constraints on what can you actually produce, you know, the constraints of your materials, constraints of time and money, money is a very big constraint in the arts. But there's not that, there's an enormous kind of like,
01:29:14
Speaker
freedom to make connections without being self-critical, without imagining a reviewer saying, no, no, no, no, you did that, but you didn't pre-register that idea and so on. So I find it very, very helpful to kind of switch between these two modes of activity. Every time I'm feeling a bit frustrated or feel like I need some new ideas, I will go and work on a drawing or
01:29:44
Speaker
work on an album, or think about a new crazy sound installation. Yeah, I think one can look at your experiments and maybe see that, right? Because they seem very playful, to use that word again, and very artistic. You could say the Baboon experiment was some kind of crazy piece of performance art.
01:30:13
Speaker
by all these baboons who didn't realize they were doing performance art, but they also didn't realize they were doing a cultural, you know, iterated learning experiment. Well, I'm very, it makes me happy to hear you say that, because that is definitely, that means I'm at least some of the steps along the way to my secret goal, not knowing the difference. I do think that I'm, I guess I am proud of
01:30:42
Speaker
whenever I feel like I've come up with a design for an experiment that feels... I suppose you kind of know when you've got a good design when it's...
01:30:55
Speaker
It feels really novel and creative and yet has a sort of sense of inevitability about it. Like, this is how it should be. Yeah, this is right. This has some kind of, this makes sense. This gets an idea across. This could communicate if it worked to the reader, the idea that you're trying to get across. And you know, that's exactly what it feels like making an artwork.
01:31:22
Speaker
Like you go, you know, I've made something like, so I do these kind of robot drawings and I still kind of struggle with it. But occasionally I'll do something and I'll look at it and go, you know, I've not seen something like that before, but it has that kind of sense of inevitability about it. It has some kind of sense that, I don't know whether it was kind of discovered rather than crafted.
01:31:52
Speaker
Yeah, it feels very, very similar. So I think that the process there is the same, ultimately. Yeah, that kind of uncanny feeling that it's almost like your agents again, when they see that thing that maps onto the world. They see that, and you get that feeling, yeah, I'm doing something right here.
01:32:17
Speaker
Yeah, I suppose we have some examples that science is just as much part of culture and it has exactly the same kind of dynamics and we're all just kind of cogs in this computational engine of culture. Well, I think that's a very big point to end on.
01:32:42
Speaker
Podcasts are part of that process too. That's right, yeah. Well, thank you so much, Simon. This has been really fun. What a wonderful tour that you've given us. Thank you, and thanks for inviting me on. It's always really nice to take a step back and try and look at the landscape from a bit further back, so thanks.
01:33:27
Speaker
So,